Towards better urban intelligence: representation learning techniques for geospatial data analytics

Abstract

With the rapid development of geopositioning and sensing technologies, urban spaces are being digitalized at an unprecedented speed. The digitalization process makes geospatial entities easily available through mobile devices. Specifically, geospatial entities refer to objects associated with geographical coordinates. They include points (e.g., points-of-interest), lines (e.g., road network), and polygons (e.g., geographical grids). The geospatial entities are relatively stable and serve as basic building blocks for location-based services. Furthermore, large amount of heterogeneous data are being generated on different geospatial entities, including mobility records, traffic sensor readings, GPS trajectories, commuting flow data, etc. These data sources attached to geospatial entities provide us great opportunities to apply data-driven techniques to understand urban dynamics, thus enhancing the urban intelligence and improving the quality of every individual’s life. In this dissertation, we aim to conduct data analytics on two types of representative geospatial entities, i.e., points-of-interest (POI) and road network, with different data sources, to reveal their underlying properties and distill knowledge for various applications. Specifically, we propose deep learning models to tackle four research problems: (1) joint mobility and time prediction on POI; (2) mobility recovery on POI; (3) semantic relationship inference among POI; and (4) representation learning for road network. With our developed methods, we can gain more insights on geospatial entities and develop more effective location-based services.